DualNetM: an adaptive dual network framework for inferring functional-oriented markers.

Abstract

Background: Understanding how genes regulate each other in cells is crucial for determining cell identity and development, and single-cell sequencing technologies facilitate such research through gene regulatory networks (GRNs). However, identifying important marker genes within these complex networks remains difficult.

Results: Consequently, we present DualNetM, a deep generative model with a dual-network framework for inferring functional-oriented markers. It employs graph neural networks with adaptive attention mechanisms to construct GRNs from single-cell data. Functional-oriented markers are identified from bidirectional co-regulatory networks through the integration of gene co-expression networks. Benchmark tests highlighted the superior performance of DualNetM in constructing GRNs, along with a stronger association with biological functions in marker inference. In the melanoma dataset, DualNetM successfully inferred novel malignant markers, and survival analysis results showed that multiple novel markers were associated with lethality in malignant melanoma. Additionally, DualNetM identified stage-specific functional markers and clarified their specific roles in mouse embryonic fibroblast reprogramming. DualNetM's marker inference function demonstrated stronger biological relevance during primed reprogramming.

Conclusions: In summary, DualNetM effectively facilitated the inference of functional-oriented markers from complex GRNs.